Merge branch 'v0.15.1-dev-wangaq' into 'v0.15.1-dev'

use tunning w4a16 moe

See merge request dcutoolkit/deeplearing/vllm!535

vllm/model_executor/layers/fused_moe/fused_moe.py

@@ -36,6 +36,7 @@ try:
     per_token_group_quant_int8, per_token_quant_int8)
     from lmslim.layers.fused_moe.fuse_moe_int8 import (fused_experts_impl_int8, get_w8a8moe_json)
     from lmslim.layers.fused_moe.fuse_moe_w4a8 import fused_experts_impl_w4a8
+    from lmslim.layers.fused_moe.fuse_moe_w4a16 import get_moe_triton_config_w4a16
 except Exception:
     print("INFO: Please install lmslim if you want to infer the quantitative model of moe.\n") 
     
@@ -1984,7 +1985,15 @@ def fused_experts_impl(
         use_nn_moe=use_nn_moe,
     )
 
-    config = get_config_func(M)
+    # config = get_config_func(M)
+    _, N1, _ = w1.shape
+    _, N2, K2 = w2.shape
+    config, _, status = get_moe_triton_config_w4a16(
+        M, E, N1, N2, K2 * 2, top_k_num, block_shape[1], hidden_states.dtype
+    )
+    # debug
+    # print(f"M:{M}, E:{E}, N1:{N1}, N2:{N2}, K2:{K2}, top_k_num:{top_k_num}, block_shape:{block_shape}, dtype:{hidden_states.dtype}, status:{status}")
+    assert status, "config not found."
 
     # We can reuse the memory between these because by the time we need
     # cache3, we're done with cache1